Onchain Agents
A scored index of 486 tools for building onchain AI agents — frameworks, MCP servers, wallets, payment rails, data feeds and trading venues, with activity and scores refreshed every day.
Use it in one command. Add it to Claude Code as an MCP server (Streamable HTTP, no key, no account):
claude mcp add --transport http satohub https://satohub.ai/api/mcp
Any other MCP client: point it at https://satohub.ai/api/mcp (46 tools, 42 of them read-only; other clients). No MCP client: curl -s https://satohub.ai/api/export/index.json returns the whole catalog as JSON.
- What this is. 486 listings of what onchain AI agents are built from, across 28 chains, rendered daily from a public export. Discovery never lists anything on its own: a project is listed only after it passes an evidence check. Nothing in this repo can be bought — listing, order and score follow the same rules for every project (NEUTRALITY.md).
- Token. Sato Hub has no token. A token using the Sato name is not ours.
- What a Sato Score is. A 0–100 measure of how open, active and verifiable a project is, computed from evidence only. It is not a safety, quality, security or returns grade, and self-reported is never treated as verified.
- How current it is. Re-rendered every day;
Last activityis observed, not claimed.unknownmeans we could not measure it, never zero. - License. Catalog data CC-BY-4.0: reuse it anywhere, credit data by satohub.ai. Tooling MIT © Prime Signal LLC. JSON, CSV and NDJSON exports plus the MCP endpoint above — no key, no account.
Runnable agent packages
Whole agents, not components: each is a sealed, signed package you download and run on your own machine. The task is the package's own statement; the evidence label is what Sato Hub recorded, word for word. Agents find them with the MCP tools onchain_agent_search_listings and onchain_agent_inspect_listing.
Verify the download against the published key before you run it: verify-release.mjs (node verify-release.mjs <package.zip> manifest.json).
The stack, layer by layer
Nine layers, from the runtime down to the venue. Each table is the highest-scored listings in that layer — highest-scored is not "best", and a score measures openness, activity and verifiability, nothing else. Checked says what we did: reproduced the documented install in an isolated container, or probed the live endpoint.
Agent runtimes & frameworks
The loop that runs the agent: prompting, memory, tool calls, scheduling. Pick this first — it decides what everything else plugs into.
Highest-scored in this layer, 6 of 30. All 30 →
How to choose. Score measures openness and activity, not adoption. A closed-source framework scores low because its code cannot be read, not because it is worse at running agents.
Onchain action kits & SDKs
Libraries that turn an agent's decision into a signed transaction — swap, transfer, mint, stake — without writing the chain plumbing yourself.
Highest-scored in this layer, 6 of 70. All 70 →
How to choose. Check the Chains column before the score: an action kit that does not cover your chain is the wrong kit at any score.
MCP servers
Tool servers your agent calls over the Model Context Protocol. This is the largest layer in the index and the fastest-moving.
Highest-scored in this layer, 8 of 220. All 220 →
How to choose. An MCP server is a remote process you hand your agent's tool calls to. Read what it asks for — a key, a signer, shell access — before you connect it.
Wallets, keys & permissions
Where the agent's key lives and what it is allowed to do: embedded wallets, MPC, smart accounts, session keys and spend permissions.
Highest-scored in this layer, 6 of 35. All 35 →
How to choose. The score says nothing about custody. Read the open-source status and the project's own docs on who can move funds; this index does not audit key handling.
Data, RPC & indexing
What the agent knows before it acts: RPC access, market data, chain indexing, wallet and token analytics.
Highest-scored in this layer, 6 of 80. All 80 →
Payments & agent commerce
How agents pay and get paid: x402 (HTTP 402 plus stablecoin), agent commerce protocols, and the stablecoin rails underneath.
Highest-scored in this layer, 6 of 148. All 148 →
How to choose. x402 support in this index means the standard is declared. Whether an endpoint actually answers HTTP 402 is measured separately — see Numbers this week.
Identity, discovery & standards
How an agent is identified and found by other agents: ERC-8004 registration, MCP discovery, A2A agent cards.
Highest-scored in this layer, 6 of 46. All 46 →
How to choose. Registering an identity proves a key controls a record. It is not a claim about the agent behind it, and nothing in this layer verifies behaviour.
Security & preflight
Checks you run before an agent installs, connects, signs or trades: contract and token screening, endpoint checks, monitoring.
Highest-scored in this layer, 6 of 17. All 17 →
How to choose. Nothing in this layer makes an agent safe. These tools surface signals; the decision, and the loss, stay yours.
Trading & DeFi venues
Where the agent actually trades: DEXs, aggregators, perps, lending and yield venues with programmatic access.
Highest-scored in this layer, 6 of 60. All 60 →
How to choose. No performance, returns or profitability is measured here, and none is implied by a score.
Starting stacks
Five ways the layers fit together, each built only from listings in this index and picked by the highest Sato Score in each layer. These are examples, not endorsements — we have not run these combinations end to end, and a higher score does not mean a better fit for your problem.
- Trade on Base — An agent that watches a market and swaps on Base.
Runtime ElizaOS → Actions Coinbase AgentKit → Wallet Bankr → Data CoinGecko MCP → Venue CloddsBot - Trade on Solana — The same shape, on Solana.
Runtime ElizaOS → Actions Coinbase AgentKit → Wallet Bankr → Data CoinGecko MCP → Venue Raydium - Get paid per call — An agent that sells a service and settles over x402.
Runtime OpenClaw → Tools Hermes Agent → Wallet Bankr → Payments Coinbase AgentKit → Identity
Per-layer pages with every listing: docs/index.md · per-listing pages: docs/listings/ · the full index is below.
Every product listing carries a Sato Score — a 0–100 measure of how open, active, and verifiable it is. Independently checked means we reproduced its documented install in an isolated container, reviewed its verification evidence, or probed its live endpoint ourselves.
🔌 Query it from your agent — one click:
Claude Desktop: paste https://satohub.ai/api/mcp into Settings → Connectors. Claude Code: claude mcp add --transport http satohub https://satohub.ai/api/mcp. Everything else → MCP quickstart
🛠️ Build your agent — Agent Architect · describe it in a sentence, get the stack → satohub.ai/build
🖥️ SATO OS — Onchain Agent Mission Control · run it on any chain, any model, your machine
The stack | Starting stacks | Index | ⬡ Top | 📈 Movers | Live Agents | Standards | New this week | Contribute | satohub.ai ↗
Built for agents as much as humans: a stable data/index.json (versioned schema), data/index.csv, llms.txt, and a live MCP endpoint your agent can query directly.
curl -s https://satohub.ai/api/export/index.json | jq '.resources[0]'
Use the data
No key, no account, no rate-limit deal to sign. A Sato Score is a measure of how open, active and verifiable a project is — not a safety, quality or returns grade.
Swap through a gate
An agent that can trade is an agent that can be talked into trading. Sato Swap
is the other answer: the agent never holds a key, and what it asks for is an
unsigned transaction. One call checks four targets — both tokens, the venue
endpoint and the recipient — applies the owner's policy server-side, simulates
the route, and returns either an unsigned object or the rule under which one was
withheld. unknown refuses by default, and every refusal names the rule, the
limit and the value it read. Nothing in the path holds a key, signs or
broadcasts. The method · full rules.
Legend
- ⬡ Score — the Sato Score: a 0–100 evidence-based measure of how open, active, and verifiable a project is. It is not a safety, quality, or returns grade (NEUTRALITY.md). Click a score for the full report.
- Activity — observed liveness: 🟢 Active (≤30d) · 🟡 Recent (≤90d) · 🟠 Quiet (≤1y) · ⚪ Dormant.
- ★ — GitHub stars (refreshed daily).
- ✓ — install reproduced by our Docker verification harness (methodology).
- Every score links to the project's page on satohub.ai — full detail, provenance, and score history live there.
- Want the stack picked for you? The Builder turns "what my agent should do" into a grounded plan from this index.
Index
- Agent Frameworks (30)
- MCP Servers (143)
- Wallet Infrastructure (35)
- Trading & DeFi (60)
- Data & APIs (83)
- Developer Tools (39)
- Skills & Plugins (31)
- Launchpads & Marketplaces (13)
- Security (17)
- Live Onchain Agents (directory) (6)
- Research & Reference (3)
- Other (2)
- Uncategorized (24)
- Runnable agent packages
- The stack, layer by layer
- Starting stacks
- Use the data
- ⬡ Top of the index
- 📈 Movers (7d)
- 📊 Numbers this week
- Live agents (Passport registry)
- Standards & protocols
- New this week
- Community
- Support the index
Openness, activity, and verifiability — not a safety grade.
30 entries · build & run agents — onchain-native frameworks and general harnesses · Related: Skills & Plugins · Developer Tools · Browse + filter on satohub.ai →
143 entries · plug-in tool servers your agent calls over the Model Context Protocol · Related: Skills & Plugins · Data & APIs · Browse + filter on satohub.ai →